Burn severity classifications derived from multitemporal Landsat Thematic Mapper images and the Normalised Burn Ratio (NBR) are commonly used to assess the post-fire ecological effects of wildfires. Ongoing efforts to retrospectively map historical burn severity require defensible, objective methods of classifying continuous differenced Normalised Burn Ratio (dNBR) data where field data are often unavailable. For three fires, we compare three methods of classifying pre- and post-fire Landsat data: (1) dNBR classification using Composite Burn Index (CBI) field data to assign severity classes; (2) fuzzy C-means classification of a dNBR image; (3) local Getis Ord statistic (Gi*) output applied to a dNBR image, classified using fuzzy C-means clustering. We then use a Kappa statistic to evaluate the agreement of severity classes assigned to a pixel with its corresponding CBI plot. For two of the three fires, the C-means clustering of the dNBR and the Gi* output performed as well or better than dNBR images classified using CBI data, with strong agreement for moderate- and high-severity classes. These results suggest that clustering of dNBR data may be a suitable approach for classifying burn severity data without field data. This method may also be useful as a tool for rapid post-fire assessments (e.g. Burned Area Emergency Response and Burned Area Reflectance Classification maps), where images must often be classified quickly and subjectively. Further analysis using additional field data and across different vegetation types will be necessary to better understand the importance of localised spatial variability in classifying burn severity data or other remote sensing change-detection analyses.